Tuesday, May 31, 2016

Indiscriminate use of CTEs considered harmful

Common Table Expressions are a wonderful thing. Not only are they indespensible for creating recursive queries, but they can be a powerful tool in creating complex queries that are comprehensible. It's very easy to get lost in a fog of sub-sub-sub-queries, so using CTEs as a building block can make things a lot nicer.

However, there is one aspect of the current implementation of CTEs that should make you pause. Currently CTEs are in effect materialized before they can be used. That is, Postgres runs the query and stashes the data in a temporary store before it can be used in the larger query. There are a number of consequences of this.

First, this can be a good thing. I have on a number of occasions used this fact to good effect to get around problems with poorly performing query plans. It's more or less the same effect as putting "offset 0" on a subquery.

However, it can also result in some very inefficient plans. In particular, if CTEs return a lot of rows they can result in some spectacularly poorly performing plans at the point where you come to use them. Note also that any indexes on the underlying tables will be of no help to you at all at this point, since you are no longer querying against those tables but against the intermediate result mentioned above, which has no indexes at all.

This was brought home to me forcefully on Friday and Saturday when I was looking at a very poorly performing query. After some analysis and testing, the simple act of inlining two CTEs in the query in question resulted in the query running in 4% of the time it had previously taken. Indiscriminate use of CTEs had made the performance of this query 25 times worse.

So the moral is: be careful in using CTEs. They are not just a convenient tool for abstracting away subqueries.

There has been some discussion about removing this aspect of the implementation of CTEs. It's not something that is inherent in CTEs, it's simply a part of the way we have implemented them in PostgreSQL. However, for now, you need to be familiar with the optimization effects when using them, or you might face the same problem I was dealing with above.